May 10, 2026

Martin Alderson published a post arguing that open-weight models are quietly closing up, and that the economic consequences could be significant.

I run, on daily basis, open-weight models locally with Ollama for two distinct reasons: first: development and testing, where local inference is fast and free; second:client work where data can’t leave the premises.

The second use case is the one that doesn’t have a clean fallback. A frontier API isn’t a substitute when the constraint is data residency, not capability.

Alderson’s argument is that open weights have functioned as a price ceiling on frontier APIs: the threat of a viable alternative keeps prices from going too high, even if most users don’t actually switch. And the recent changes he lists (Meta removing weights from Muse Spark, Alibaba going API-first, Kimi K2.6 adding attribution clauses, Mistral adding commercial restrictions) suggest that ceiling may be eroding.

Most of the meaningful open-weight releases in the last year have come from Chinese labs, where the incentives are different. That’s not a stable foundation for an assumption as load-bearing as “there will always be a capable open alternative.”